AI Wiki · Mediocre Security
Concept

Probe (interpretability)

A small detector that reads a model's internal signals at each step to flag risky behavior, rather than judging only its outputs.

An interpretability probe is a lightweight classifier trained on a model’s internal activations to detect a property, such as “this step is part of a hacking attempt”, directly from what the model is computing rather than from its text. Goodfire’s monitors run probes at each agent step and escalate to a second model only when flagged; OpenAI’s staged misalignment monitoring also starts with internal-signal checks. Probes matter because a model’s stated reasoning may not reflect what it is doing (see GPT-6 Astra).

Connections

Probe (interpretability) and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →

What links here

Concept Alignment

Company Goodfire